DeepSeek has released an experimental V4 Flash Vision, a version of its flagship model that can now work not only with text, but also with images and screenshots. The company claims that in tests of AI agents working with visual information, the model is already approaching Anthropic's Claude Opus 4.8.
DeepSeek has released an experimental V4 Flash Vision, a version of its flagship model that can now work not only with text, but also with images and screenshots. The company claims that in tests of AI agents working with visual information, the model is already approaching Anthropic's Claude Opus 4.8.
This was reported by DeepSeek.
The regular DeepSeek V4 Flash only works with text. But the new DeepSeek-V4-Flash-Vision-Exp has gained “sight.” Now you can show it a screenshot of a website, a graph, a document, or another image and ask it to figure out what’s happening there.
However, these results were published by DeepSeek itself, and there is no independent testing of the V4 Flash Vision yet. Therefore, how close the model will be to the Claude Opus 4.8 in third-party benchmarks and real-world tasks remains to be seen.
The peculiarity of this model is how many tokens an image takes up. The company has set a limit of 384 tokens per image, and they are paid for at the usual V4 Flash tariff. For comparison, in other models the number of tokens often increases with the size or detail of the image. For example, Anthropic indicates that a 1000×1000 pixel image takes up 1296 tokens in Claude. And Gemini breaks large images into separate fragments of 258 tokens each.
However, DeepSeek does not feed the model a large image at full size, but first reduces it to approximately 800×800 pixels. However, it is not yet clear whether small text, buttons, and other interface details are lost when this reduction is made.
Previously, dev.ua wrote that DeepSeek had reduced the cost of the V4 Pro model by 75%.
